Download Distributed Uplink Power Control for Optimal SIR Assignment in

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Marginalism wikipedia , lookup

Marginal utility wikipedia , lookup

Choice modelling wikipedia , lookup

Microeconomics wikipedia , lookup

Transcript
1
Distributed Uplink Power Control for Optimal SIR
Assignment in Cellular Data Networks
Prashanth Hande1,2 , Sundeep Rangan2 , Mung Chiang1 , Xinzhou Wu2
1
Department of Electrical Engineering, Princeton University, NJ 08544, USA
2
Qualcomm Flarion Technologies, NJ 07921, USA
Abstract—This paper solves the joint power control and SIR assignment problem through distributed algorithms in the uplink of
multi-cellular wireless networks. The 1993 Foschini-Miljanic distributed power control can attain a given fixed and feasible SIR
target. However, feasibility of SIR target cannot be readily determined a priori, and, in addition, SIR needs to be jointly optimized
with transmit powers in wireless data networks. In the vast research literature since the mid-1990s, solutions to this joint optimization problem are either distributed but suboptimal, or optimal but centralized. For convex formulations of this problem, we
report the first distributed and optimal algorithm.
The main issue that has been the research bottleneck for many
years is the complicated, coupled constraint set, and we resolve it
through a re-parametrization via the left Perron Frobenius eigenvectors, followed by development of a locally computable ascent direction. A key step is a new characterization of the feasible SIR
region in terms of the loads on the base stations, and an indication
of the potential interference from mobile stations, which we term
spillage. Based on this load-spillage characterization, we first develop a distributed algorithm that can achieve any Pareto-optimal
SIR assignment, then a distributed algorithm that picks out a particular Pareto-optimal SIR assignment and the associated powers through utility maximization. Extensions to power-constrained
and interference-constrained cases are carried out. The algorithms
are theoretically sound and practically implementable: we present
convergence and optimality proofs as well as simulations using
3GPP uplink evaluation tools.
I. I NTRODUCTION
Power control in cellular networks has been extensively studied since the late 1980s as an important mechanism to control Signal-to-Interference Ratios (SIR), which in turn determine Quality-of-Service (QoS) metrics such as rate, outage, and
delay. Uplink power control over multiple cells is particularly
challenging. The basic version of CDMA power control solves
the near-far problem by equalizing the received powers from
mobile stations (MS) located at different parts of the cells. More
sophisticated, iterative power control algorithms invented since
the early 1990s find a transmit power vector so as to ensure that
each MS attains the target SIR [2] while the overall power consumption is minimized. In particular, a distributed algorithm to
achieve given fixed targets of SIR , if it is feasible, was proposed
by Foschini and Miljanic [3] in 1993, with wide applicability in
current cellular networks. This result has been followed by a
This work was supported NSF Grants CCF-0448012, CNS-0417607, and
CNS-0427677. Part of this work was presented at IEEE INFOCOM 2006.
large body of publications. In particular, the convergence properties of a more general class of similar distributed algorithms
for power control was studied in [4].
However, there are two major limitations of the Foschini Miljanic power control algorithm. First, while in the power control
algorithm in [3], the target SIRs are assumed to be within the
feasibility region, a distributed mechanism to check the feasibility of the SIRs remains an open problem. A second limitation
is that the fixed SIR approach to the power control problem is
suitable only for voice networks where a given minimum SIR
is necessary for stable communication at each MS. A cellular
data network, however, can assign SIR according to traffic requirements, with higher SIRs implying better data rates and possibly greater reliability, while smaller SIRs can provide lower
data rates for non-real-time data services. In addition, a cellular
operator of a wireless data network might want to treat higher
tariff paying users preferentially by allocating them to higher
QoS classes, and the SIR target can be set differently for users
in different QoS classes. It is therefore important to find a distributed mechanism to assign SIRs that belong to the feasible
set and yet optimal according to criteria defined by the network
operator. We provide a solution to this problem in this paper.
A main reason that such a solution has not been found over
the last 16 years is that the feasible SIR region is coupled in a
complicated way across MS in different cells. A standard way
to describe the feasibility region is in terms of a condition on the
spectral radius of a matrix with network-wide parameters as defined in 1992 [2], which seemingly requires a centralized computation. Initial attempts to assure some QoS performance but
maintain the realistic distributed computation mechanism were
made in the mid-1990s, e.g., in [5], [6]. Since the late 1990s,
there have been two main threads of work tackling the problem
of jointly optimizing SIR assignment and transmit powers over
the feasibility region, e.g., in [7], [8], [9], [10], [11]. First, the
work in [7] in 2001 proposes a utility framework by defining
utility functions of the rates assigned to the MSs, with the rates
being functions of SIRs, and the distributed solution converges
to a Nash equilibrium that may be socially suboptimal. Second,
in [12] in 2004, it is shown that the feasible SIR region is convex in the logarithm of SIR. This makes it possible to propose
the utility framework for power control as a convex optimization
problem. This has been the approach taken in [9], [10], [11] to
obtain the global optimum, but the solutions require complex
and centralized computation.
2
The main thread of this paper is the development of distributed algorithms that solves the convex formulations of the
problem of jointly optimizing powers and SIR assignment over
the feasibility region. Pictorially, the problem is to find the
utility-optimal SIR assignment (where contours of the utility
function are shown in dotted lines in Figure I) over the feasible
SIR region (shown as the shaded region). The core idea behind
all the developments and proofs is a novel re-parametrization
of the feasibility region through the left, rather than the right,
Perron-Frobenius eigenvectors of a network parameter matrix.
Mathematically it amounts to a nonlinear change of coordinate
that leads to decoupling through dual decomposition. The underlying engineering intuition is that the quality of a wireless
uplink depends on both the strength of its direct channel gain
and the weakness of interfering channel gains. The second key
step in the solution, after the re-parametrization, is the development of an ascent direction that can be computed locally at each
MS without global coordination.
12
10
SIR 2
8
6
4
2
0
0
1
2
3
4
5
6
7
8
9
10
SIR 1
Fig. 1. Utility maximization over coupled feasibility region.
The key results of this paper are summarized as follows:
• Characterize the Pareto-optimal boundaries of feasibility
regions. We define three different SIR feasibility regions.
The first is a ρ-optimal subset where the power and interference are finite by a suitable choice of the parameter ρ.
The other two feasibility sets are defined by imposing explicit constraints on the maximum allowed power and interference, respectively. We then derive the Pareto-optimal
boundary of each of the three sets, with the intention of
operating on those boundaries for SIR assignment.
• Re-parameterize feasibility regions. We propose an alternative view of the system, different from the standard one
of power and interference. This is done by introducing two
sets of parameters that can be interpreted as the load on
the network, and an indication of the potential emitted interference from mobile stations, which we term spillage.
It is of practical importance to note that the terms of load
and spillage are used in a loose way in wireless industry.
Here they have precise mathematical definitions, and the
intuitions behind them are used in a provably optimal way.
• Distributed algorithms to attain any point on the Paretooptimal boundary. Utilizing this load-spillage characterization, we propose a distributed algorithm that can attain
any point on the Pareto-optimal boundary.
Distributed algorithms to “slide” along the Pareto-optimal
boundary towards the utility-optimal point. Further optimization over the Pareto-optimal boundary for unique operating point is posed as a utility maximization problem
to maximize network utility. A distributed solution is proposed (the main algorithm labeled as Algorithm 4), thus
overcoming the bottleneck of centralized computation of
Perron-Frobenius eigenvalues and eigenvectors as required
in current literature.
• In the process, we also introduce a new utility function,
which we term the pseudo-linear utility function. It approximates the linear utility function at high SIRs but
avoids MS starvation at low SIRs by precluding zero SIR
allocation.
• Our algorithms can be readily implemented in today’s cellular networks. Performance in realistic, large-sale networks is evaluated using the 3GPP uplink evaluation tool.
The rest of this paper is organized as follows. In Section IIA, we discuss the system setup and feasible SIR region, then
introduce three different feasibility sets based on practical constraints, and then define the Pareto-optimal boundaries of the
three sets. We develop the load-spillage characterization of
the system in Section III-A, based on which we propose distributed algorithms of joint SIR assignment and power control
for Pareto-optimal configuration in Section III. We discuss the
utility framework in Section IV and propose distributed algorithms for utility optimization in Section V. Convergence and
global optimality proofs are stated in the main text, with proofs
of key lemmas summarized in various Appendices. The algorithm is extended to the power constrained and interference
cases in Section V-B. Finally, we present simulations demonstrating the performance and robustness of our algorithms in
Section VI. We conclude the paper in Section VII with a discussion of future research: the integration of multi-user detection,
opportunistic scheduling, bandwidth assignment, soft handoff,
and antenna beamforming into the load-spillage framework in
this paper.
•
II. F EASIBILITY
AND
PARETO -O PTIMALITY
A. System Setup
Consider a general multi-cell setup where M MSs establish
links to N BSs, as illustrated in Fig 2. We assume that each MS
is served by one of the N BSs, thereby establishing M links.
We let σi denote the receiving BS for link i.
Let Ci denote the set of links whose transmitted power appears as interference to link i. This definition allows us to consider both orthogonal and non-orthogonal uplinks. In a nonorthogonal uplink, such as CDMA, transmitted power from all
links appear as interference, so we set Ci = {j | j 6= i}. For
an orthogonal uplink, such as OFDM, links terminating on the
same BS are orthogonal and do not contribute interference to
one another. In this case, we set Ci = {j | σj 6= σi }.
Now let h0kj denote the absolute path gain from MS j to BS
3
Let γi be the SIR achieved by link i. With the above notation,
γi = pi /qi , or equivalently,
Interference
BS1
Power
p = D(γ)q,
(5)
where D(γ) = diag(γ1 , . . . , γM ). Combining (4) and (5), we
get the following basic equations:
BS0
MS1
MS0
q = GD(γ)q + η,
(6)
p = D(γ)Gp + D(γ)η.
(7)
and
BS2
B. SIR Feasibility Region
Fig. 2. An example of a multi-cellular network.
k, and define the normalized path gain by
hkj = h0kj /h0σj j .
Define the M × M matrix G by,
hσi j if j ∈ Ci ,
Gij =
0
if j 6∈ Ci ,
(1)
which represents the normalized path gain from MS on link j
to the receiving BS on link i, when link j interferes with link
i. Since an orthogonal uplink has Ci = {j | σj 6= σi }, and a
non-orthogonal uplink has Ci = {j | j 6= i}, we can write both
cases using a single expression:

 hσi j if σi 6= σj
Gij =
0
if i = j
(2)

θ
if σi = σj , i 6= j
where θ = 1 for the non-orthogonal case, and θ = 0 for the
orthogonal case.
Let pj be the received signal power on link j at its serving BS
σj . Since h0σj j is the path gain from MS on link j to its serving
BS, the MS on link j must transmit at a power of pj /h0σj j . At
any BS k, this signal will appear with a power of
h0kj pj /h0σj j = hkj pj .
If j ∈ Ci , this transmission will appear as interference to link i
with a power of hσi j pj = Gij pj . Using the fact that Gij = 0
for j 6∈ Ci , the total interference and noise at the BS serving MS
i is given by
qi =
X
j∈Ci
Gij pj + ηi =
M
X
Gij pj + ηi ,
(3)
j=1
where ηi ≥ 0 is the power of noise other than interference from
other links. In matrix notation, (3) can be written as,
q = Gp + η.
(4)
Due to the interference between links, not all SIR vectors
γ are achievable. An SIR vector γ ≻ 0 is called feasible if
there exists an interference vector, q 0, and power vector
p 0, satisfying (6) and (7), respectively. Let ρ(·) denote the
spectral radius function 1 and assume that G is a primitive matrix [16] 2 . The following standard result gives a simple spectral
radius characterization of SIR feasibility.
Lemma 1—Zander [2]: An SIR vector γ ≻ 0 is feasible if
and only if
1) ρ(GD(γ)) < 1, when η 6= 0
2) ρ(GD(γ)) = 1, when η = 0.
If η 6= 0, then the power allocation pertaining to γ is denoted
by p(γ) and uniquely given by
p(γ) = (I − D(γ)G)−1 D(γ)η.
(8)
Correspondingly, the interference vector is denoted by q(γ) and
uniquely given by
−1
q(γ) = (I − GD(γ))
η.
(9)
If η = 0, then the power allocation p(γ) is the right eigenvector of GD(γ) and the interference vector q(γ) is the right
eigenvector of D(γ)G, both corresponding to eigenvalue 1. The
eigenvectors are unique up to a scaling factor.
For a given feasible SIR allocation, the Foschini-Miljanic
power control algorithm in [3] converges to the unique power
vector as given in equation (8) when η 6= 0 and to the right
eigenvector of GD(γ) when η = 0. If γi is the target SIR
and γ̂i is the measured SIR at MS i, then the transmit power or
effectively the received power is updated in the algorithm as
pi [t + 1] = pi [t]γi /γ̂i .
(10)
However, the power control algorithm assumes a given SIR target without trying to jointly optimize over SIR assignments and
transmit powers. Furthermore, the given SIR target is assumed
to be feasible, but distributively checking feasibility is a difficult
task on its own.
1 Spectral radius is the maximum of the absolute value of the eigen values of
a matrix.
2 This is equivalent to saying that the network represented by G is connected,
a reasonable assumption.
4
We assume in the rest of this paper that η 6= 0 unless stated
otherwise. Let B = {γ ≻ 0 : ρ(GD(γ)) < 1} denote the
set of all feasible SIR vectors γ. As γ approaches the boundary
of the feasibility region, the interference and power vectors tend
to infinity in general. Practical networks, however, have finite
limits on the interference and received power. To take care of
this technical point, we devote several paragraphs to describe
three alternative ways to ensure finite powers in practice.
Let qm ≻ 0 and pm ≻ 0 be maximum interference and
power vectors, and define the corresponding feasibility sets of
SIR vectors,
B(pm ) =
B(qm ) =
{γ ∈ B | p(γ) pm },
{γ ∈ B | q(γ) qm }.
The set B(pm ) limits the received power, pi of each mobile
i. By scaling the received power by the path gain from the mobile to its serving BS, this bound can incorporate transmit power
limits. The bound is most useful in coverage limited networks
where the transmit power capabilities of the mobiles are the limiting factor in network capacity.
In capacity-limited networks, however, cells are typically sufficiently densely deployed, so that the absolute power capabilities of the mobiles are not the limiting factor. The set B(qm )
limits the interference qi at the BS serving MS i. In commercial
network specifications, the interference limit qm is often stated
in the form qm = κη for some constant κ ≥ 1. With this definition, the interference, qi , at each BS i, is not allowed to be larger
than a factor κ greater than the thermal noise ηi . The factor, κ, is
called the rise over thermal (ROT), and typically quoted in dB,
ROT = 10 log10 (κ).
The ROT limit bounds the additional interference to the cell,
and thereby limits the power required for new mobiles to access
the network. Typical ROT values in commercial networks range
from 3 to 10 dB and we will employ similar constraints in the
simulations in Section VI.
Another constraint set that we will consider is
Bρ = {γ ≻ 0 | ρ(GD(γ)) ≤ ρ},
which is defined for any ρ ∈ [0, 1) 3 . By Lemma 1, any γ ∈ Bρ
is feasible. The constraint, γ ∈ Bρ , results in finite power and
interference for ρ < 1.
The sets B, B(pm ), B(qm ) and Bρ are not, in general, convex. However, a result in [12] shows that a certain transformation of the sets are. Given a set of feasible SIR vectors Γ ⊆ B,
let log Γ = {log γ | γ ∈ Γ}.
Lemma 2—Boche and Stanczak [12]: The functions q(γ),
p(γ) and ρ(GD(γ)) are convex in log γ in the region log B.
In particular, the sets, log B(qm ), log B(qm ) and log Bρ are
convex.
3 We tolerate the abuse of notation where ρ is both the spectral radius function
as well as a parameter that limits the spectral radius.
C. SIR Pareto-Optimality Boundary
Let Γ ⊆ B be a set of feasible SIR vectors. The selection of
a feasible γ ∈ Γ is, in general, a multi-objective optimization
problem. Increasing the SIR, γi , for one mobile may require
that the SIR, γj , for another mobile be reduced. A feasible SIR
vector γ ≻ 0 is called Pareto-optimal if it is impossible to increase the SIR of any one link without at the same time reducing
the SIR of some other link. The Pareto-optimal points of a set Γ
form the Pareto-optimal boundary which we denote by ∂Γ. The
following theorem, proved in Appendices A and B, provides a
simple characterization of Pareto-optimality for the feasible sets
defined in Section II-B.
Theorem 1: Let η > 0 be any positive noise vector, and let
the gain matrix G be primitive. Then, the Pareto-optimal boundary ∂Γ of the following sets Γ is described as follows:
1) If Γ = B(qm ), γ ∈ ∂B(qm ) if and only if q(γ) qm
with at least one i such that qi (γ) = qim
2) If Γ = B(pm ), γ ∈ ∂B(pm ) if and only p(γ) pm
with at least one i such that pi (γ) = pm
i
3) If Γ = Bρ , γ ∈ ∂Bρ if and only if ρ(GD(γ)) = ρ.
A complete summary of the feasibility regions and Paretooptimal boundaries is provided in Table I.
III. D ISTRIBUTED P OWER C ONTROL FOR
PARETO -O PTIMAL SIR A SSIGNMENT
A. The Load-Spillage Characterization
Lemma 1 above shows that a candidate SIR vector γ is feasible if and only if ρ(GD(γ)) < 1. Unfortunately, this has not
been directly verifiable in a distributed manner. A direct computation of the spectral radius ρ(GD(γ)) will require a centralized
controller that knows the entire candidate SIR vector γ, as well
as the complete connection matrix G. In this section, we present
an alternative, “dual” parameterization of the feasible SIR vectors γ that is more easily computed in a distributed manner. The
parameterization is based on the following Lemma.
Lemma 3: An SIR vector γ ≻ 0 is feasible, i.e., γ ∈ B, if
and only if there exists a s ≻ 0 and ρ ∈ [0, 1) such that
sT GD(γ) = ρsT
(11)
Further, the SIR vector is ρ-optimal, i.e., γ ∈ ∂Bρ .
Proof: If γ is feasible, by Lemma 1, ρ = ρ(GD(γ)) ≤ 1.
Let s ≻ 0 be the left eigenvector associated with ρ(GD(γ)),
which is positive because GD(γ) is primitive [16].
Conversely, if some s ≻ 0 and ρ ∈ [0, 1) satisfies (11), then s
is a positive left eigenvector of the non-negative, primitive matrix GD(γ) which implies that ρ is the maximal eigenvalue
ρ = ρ(GD(γ)) [16]. Since ρ < 1, the corresponding γ is
feasible.
Lemma 3 leads to a natural parameterization of the set of feasible SIRs. Given any s ≻ 0 and ρ ∈ [0, 1), let
γ(s, ρ) = ρs/r(s),
(12)
where the division of the vectors is component-wise, and r(s) is
the following M -dimensional vector:
r(s) = GT s.
(13)
5
TABLE I
F EASIBILITY R EGIONS AND T HEIR A SSOCIATED PARETO -O PTIMAL B OUNDARIES
{η = 0}
{η 6= 0, ρ − optimal}
{η 6= 0, p(γ) pm }
{η 6= 0, q(γ) qm }
Feasibility Region:
{γ : ρ(GD(γ)) = 1}
B = {γ : ρ(GD(γ)) < ρ}
B(pm ) = {γ : 0 p(γ) pm }
B(qm ) = {γ : 0 q(γ) qm }
Observe that (11) is equivalent to r(s)T D(γ) = ρsT . Consequently, s, γ and ρ satisfy (11) if and only if γ = γ(s, ρ).
Therefore, Lemma 3 shows that γ(s, ρ) parameterizes the set of
all feasible positive SIRs. Moreover, if ρ ∈ [0, 1) is fixed, then
the set of γ(s, ρ) parameterizes the set of ρ-optimal SIRs.
Also, combining (13) with (11), we see that
ρrT = rT D(γ)G.
(14)
Therefore, r and s are left positive eigenvectors of the matrices,
D(γ)G and GD(γ), respectively. This pair of left eigenvectors, {s, r}, have an interesting interpretation: s represents the
load on the network to support an SIR γ, and r represents the
potential interference due to an SIR assignment. We call them
the load factors and spillage factors.
This interpretation is made more clear by writing down the
spillage ri associated with link i in terms of the link matrix Gij
and the load factors. From (12) and (13), it follows that
P
spillage: ri =
j Gji sj
(15)
load: si =
ri γi /ρ
The spillage factor ri represents the effect of interference due
to link i on other links in the network weighted by the loads of
each link. Link i, with an SIR γi and responsible for spillage ri ,
loads the network with si = (1/ρ)ri γi in the sense that it is less
tolerant by a factor of si to interference from other links in the
network.
B. Distributed SIR Assignment for ρ-Optimality
The characterization developed in Section III-A leads to a
simple distributed algorithm for assigning feasible SIRs. To
describe the assignment algorithm, suppose that we wish to
achieve an SIR vector γ = γ(s, ρ) for some s ≻ 0 and
ρ ∈ [0, 1). We let r = r(s), then it follows from (15) that
X
X
X
ri =
Gji sj =
Gji sj +
Gji sj
(16)
j
j6=i,σj =σi
j6=i,σj 6=σi
Using (2), we can write this as
X
X
ri = θ
sj +
hki ℓk
j6=i,j∈Sσi
(17)
k6=σi
where Sk = {j | σj = k} is the set of mobiles connected to BS
k, and
X
ℓk =
sj ,
(18)
j∈Sk
Pareto-optimal Boundary:
{γ : ρ(GD(γ)) = 1}
∂B = {γ : ρ(GD(γ)) = ρ}
∂B(pm ) = {γ : 0 p(γ) pm , ∃i : pi (γ) = pm
i }
∂B(qm ) = {γ : 0 q(γ) qm , ∃i : qi (γ) = qim }
which we term the BS-load factor associated with BS k. The
spillage can therefore be factored as
ri = riint + riext ,
(19)
where riext is the external component of the spillage due to MSs
in other cells,
X
hki ℓk ,
(20)
riext =
k6=σi
riint
and
is the internal component due to spillage from mobiles
in the same cell,
P
if the uplink is non-orthogonal,
j6=i,j∈Sσi sj
riint =
0
if the uplink is orthogonal.
(21)
This computation leads to a natural distributed SIR assignment algorithm. Fix the load factors sj allocated to the MSs,
which in turn fixes the BS-load factors ℓk . Then the following
algorithm assigns a ρ-optimal SIR.
Algorithm 1—Distributed Assignment over ∂Bρ :
Initialize: Fixed s ≻ 0 and ρ ∈ [0, 1).
1) BS k broadcasts the BS-load factor ℓk .
2) Compute the spillage factor ri according to (17).
3) Assign SIR values γi = ρsi /ri .
Stop. The resulting SIR vector γ = γ(s, ρ).
Remarks:
1) The algorithm consists of a distributed one-step computation (no iteration), and results in ρ-optimal SIR given a
set of load-factors {si }. Increasing si for some i results
in a higher γi at the expense of some other links in the
system. But a proportional increase in every si results in
no change in the SIR assignment.
2) BSs can broadcast ℓk on the downlink at a fixed power
so that MSs in the vicinity can decode the BS load-factor.
The BS load-factor transmission power can be fixed at a
value such that all MSs that can potentially interfere with
the BS can decode the load-factor transmitted. In addition, the transmission of the BS load-factor at fixed power
will enable the MS to measure the relative path gains hki
required for spillage calculation.
3) The initial load allocation can be made at the MS (MSControl). Alternatively, it can also be made at the BS
(BS-Control). BS-Control requires a mechanism to either
convey the load si to the MS, or convey riext from the
MS to the BS and then the assigned SIR from the BS to
6
the MS. BS-Control has a higher messaging overhead in
comparison to MS-Control but has the advantage of better overall system control since the load allocation can be
made based on a network-wide criterion.
Algorithm 1 is only the first step of the development. It results in a ρ-optimal point on the Pareto-optimal boundary of the
feasible SIR region. By optimizing further over s along the right
direction, the “best” point on the Pareto-optimal boundary can
be picked out, as will be discussed in the next two sections.
For ρ < 1, the resulting power and interference vectors from
Algorithm 1 are finite. We further quantify the resulting power
and interference vectors as follows.
Lemma 4: Let γ = γ(s, ρ) for some s > 0 and ρ ∈ [0, 1),
and let p = p(γ) and q = q(γ) be the corresponding power
and interference vectors. Then,
1
1
T
T
r
p−
D(γ)η = 0, s
q−
η = 0, (22)
1−ρ
1−ρ
where r = r(s) in (13).
Proof: Operating on the left by rT on both sides of (7) and
T
by s on both sides of (6) results in (22).
The following lemma is proved in Appendix C.
Lemma 5: Fix any s > 0 and for ρ ∈ [0, 1), let qρ = q(γ) be
the interference vector corresponding to γ = γ(s, ρ). Then,
(a) There exists a unique vector q̃ > 0 such that, for all ρ ∈
[0, 1), q̃ is a positive right eigenvector of GD(γ(s, ρ)),
normalized so that sT q̃ = sT η.
(b) The interference vectors qρ satisfy the following limit:
lim (1 − ρ)qρ = q̃.
ρ→1
(23)
To interpret Lemma 4, let y = sη, where the multiplication
of the vectors is component-wise. Then, the second equation in
(22) can be rewritten as
q
1
T
y
−
= 0.
η
1−ρ
Recalling from Section II-B that the ratio q/η is the ROT vector
for the system, Lemma 4 shows that a weighted average ROT is
bounded by 1/(1 − ρ). The parameter ρ is thus related to the
ROT limit. Lemma 5 extends this further and shows that, as
ρ → 1, ROT = O(1/(1 − ρ)). This fact will be used in a later
proof of convergence and optimality.
C. Distributed Control for Pareto-Optimality with Power and
Interference Constraints
While ρ-optimality with ρ < 1 guaranteed a finite power
and interference vector, an alternative operating mode is to
specify the power and interference constraints explicitly. In ρoptimization, the set of feasible SIR vectors, γ, were parameterized by the load vector s. For the interference and power
constrained case, we need to introduce a second parameter vector ν 0. The vector ν will be called the price vector, and will
play a role as Lagrange multipliers as will be seen in Section VB.
For ease of presentation, we define two cases, case P and case
Q, with power consraint pm and interference constraint qm , respectively.
Case P : {γ ∈ B(pm )}
(24)
Case Q : {γ ∈ B(qm )}
Given a load vector s ≻ 0 and price vector ν 0, define vectors
rp (s, ν), γ p (s, ν), rq (s, ν) and γ q (s, ν) as
Case P : rp (s, ν) =
γ p (s, ν) =
Case Q : rq (s, ν)
γ q (s, ν)
GT s + ν
s/rp (s, ν)
(25)
= GT (s + ν)
= s/rq (s, ν)
(26)
Notice the resemblance of the relationship (25) and (26) to that
in equation (12) and (13). This resemblance enables us to interpret s as a load-factor and r = rp (s, ν) as the spillage associated
with case P and r = rq (s, ν) as the spillage associated with case
Q. In fact, for a given link i, the relationship can be expressed
as
P
Case P : ri =
j Gji sj + νi
(27)
si = ri γi
P
Case Q : ri =
j Gji (sj + νj )
(28)
si = ri γi
The achieved SIR in each case is given by γ = γ p (s, ν) and
γ = γ q (s, ν), respectively. For a fixed s and a suitable choice
of the price vector ν, the achieved SIR γ = γ p (s, ν) lies on the
Pareto-optimal boundary ∂B(pm ) for case P and the achieved
SIR γ = γ q (s, ν) lies on the Pareto-optimal boundary ∂B(qm )
for case Q.
The prices {νi } correspond to local constraints and hence
amenable to a simple subgradient-based update:
+
Case P : νi [t + 1] = [νi [t] + δ[t](pi [t] − pm
i )]
(29)
Case Q : νi [t + 1] = [νi [t] + δ[t](qi [t] − qim )]+
(30)
where δ[t] = δ0 /t, δ0 > 0, is a suitable choice for the step
size. Based on this update, we present the following algorithm
to iteratively arrive at the appropriate price-vector.
Algorithm 2—Distributed Assignment over ∂B(pm ),∂B(qm ) :
Initialize: fixed si [0] ≻ 0, νi [0] 0.
1) Case P: Compute spillage ri [t] according to (27).
Case Q: Compute spillage ri [t] according to (28).
2) Case P: Assign SIR γi [t] for MS i according to (27).
Case Q: Assign SIR γi [t] for MS i according to (28).
3) Case P: Measure resulting power pi [t].
Case Q: Measure resulting power qi [t].
4) Case P: Update power price νi [t] according to (29).
Case Q: Update power price νi [t] according to (30).
Continue: t := t + 1.
While prices were additive to the spillage in case P, bringing
down the SIR assignment to lie on ∂B(pm ), the prices in case
Q are additive to the load and have the same effect of bringing
7
down the SIR assignment to lie on ∂B(qm ). When the prices
converge, the resulting SIR is pareto-optimal as shown in the
following theorem.
Theorem 2: If Algorithm 2 has a fixed point γ, then γ ∈
∂B(pm ) for case P and γ ∈ ∂B(qm ) for case Q.
Proof: Consider the case P. If ν is the fixed point of Algorithm 2, then it cannot be that ν = 0 because that results in a γ
requiring infinite power. So there is some i, such that νi > 0 and
for such a i, pi = pm
i , indicating that the resulting γ is indeed
Pareto-optimal. The argument is similar for case Q.
IV. U TILITY M AXIMIZATION OVER F EASIBILITY R EGIONS
In Section III, we discussed distributed mechanisms to
achieve any SIR vector γ on a Pareto-optimal boundary of the
feasible SIR region. Picking a particular point on the Paretooptimal boundary, out of an infinite number of choices, can be
accomplished by specifying a network-wide objective function,
generally denoted as the network utility U (γ). Given a set of
feasible SIR vectors, γ ∈ Γ ⊆ B, the optimal SIR over this set
is defined by γ opt = arg maxγ ∈Γ U (γ).
Let Ui (βi ) be a utility, representing the value to the overall
network of allocating link i a QoS metric βi , which has a bijective mapping to the SIR γi given by βi = β(γi ). One such QoS
metric of interest is the Shannon capacity, given by
β(γi ) = wi log2 (1 + γi W/wi )
(31)
where wi is the bandwidth allocated to each user out of a total
available bandwidth of W . We assume that wi is fixed and equal
for all users in this work. Joint optimization over SIR assignment and bandwidth allocation has also recently been studied
in [13].
A useful set of utility functions to consider is the α-fair utility
functions [14] defined on the QoS parameters βi = β(γi ):
log(βi )
if α = 1,
(32)
Ui (βi ) =
−1 1−α
(1 − α) βi
if α 6= 1
where α = 0 represents Linear Utility and α = 1 represents
Log Utility. Increasing α leads to fairer allocation in the corresponding utility maximization problem with proportional fairness achieved for (α = 1) and max-min fairness achieved as
(α → ∞).
In addition to the α-fair utilities, we consider a novel utility
function that we call the pseudo-linear utility function:
Pseudo-linear utility function is particularly suitable from the
network-opeartor’s point of view, since operators would typically like to maximize the sum throughput under normal loading
of the cells, but also ensure no outage to any MS when the cells
are heavily loaded. In this paper, instead of the linear utility
function, we use this pseudo-linear utility function.
The bijective mapping β of the QoS metric βi from the SIR
γi allows us to view the utility as a function of the SIR. For
ease of presentation, we tolerate some abuse of notation and
represent this function as Ui (γi ) instead of the more appropriate
Ui (β(γi )). We use the following notation: Ui′ (γi ) = ∂Ui /∂γi
and Ui′′ (γi ) = ∂ 2 Ui /∂ 2 γi . We will make the following assumptions on the utility functions throughout this paper.
1) The utility functions Ui (γi ) are strictly increasing, twice
differentiable and strictly concave in log γi .
2) Ui′ (γi ) and Ui′′ (γi ) are both bounded for all γi ≥ 0.
3) Ui (γi ) is fair in the sense that, as γi → 0, Ui (γi ) → −∞,
so that zero SIR assignment is precluded in any solution
to the utility maximization problem with non-empty constraint set.
The assumption on strict concavity in log γi can be expressed in
an alternative form. It can be easily verified that Ui (γi ) is strictly
concave in log γi if and only if the negative of the curvature is
sufficiently large:
Ui′′ (γi ) ≤ −
Ui′ (γi )
.
γi
In particular, the above condition is satisfied for α-fair utilities
when α ≥ 1.
Together with Lemma 2, the above assumption shows that
the maximization of the utility over any of the feasibility sets
Bρ , B(qm ), or B(pm ) is a convex optimization problem. In
particular, a local maximum of the constrained optimization is
also a global maximum.
V. D ISTRIBUTED P OWER C ONTROL FOR U TILITY-O PTIMAL
SIR A SSIGNMENT
In this section, we develop a distributed power control algorithm to solveP
the optimal SIR assignment problem of maximizing U (γ) = i Ui (γi ) when the SIR is constrained to one of
three feasibility sets: Bρ , B(pm ), B(qm ).
Intuitively, any such optimization over the Pareto-optimal
boundary should assign a higher SIR value to links in good
channel conditions. A second criterion, which is less understood
Pseudo-Linear Utility: Ui (βi ) = log(exp(βi /W ) − 1). (33) in existing literature, is to take into consideration the channel
This utility function is fairer than the linear utility function in conditions of the interfering paths of each link. Good chanan important way: zero QoS assignment is infinitely penalized nel conditions on the interfering paths can carry over substanand not allowed. At high values of the QoS metric, the pseudo- tial interfering signal strength even if the transmit power of the
linear utility approximates the linear utility, thus restoring the link is low. In this case, we should assign a higher SIR to the
capability of maximizing a close approximation to sum QoS. At link that has worse interfering channels. This intuition will be
low values of the QoS metric, the pseudo-linear utility approx- made precise in this section. The power-interference viewpoint
imates the logarithmic utility and ensures non-zero QoS for all of the system provides information on the condition of the channel that represents the primary connection of a link, whereas the
transmitting users. More precisely, for Ui (βi ) in (33),
load-spillage viewpoint of the system provides information on
Ui (βi ) → βi /W as βi → ∞
the channel conditions of the interfering paths associated with
Ui (βi ) → −∞ as βi → 0.
(34) the link.
8
A. Distributed Optimization over ρ-Optimal Boundary
First consider the case where γ is a ρ-optimal assignment for
some ρ ∈ [0, 1):
maximize U (γ)
subject to
(35)
γ ∈ Bρ .
While the above problem is convex optimization under the
aforementioned assumptions, the constraint set is highly coupled. As a result, no distributed algorithms have been possible
to date. Indeed, the problem has been considered before (e.g., in
[9], [10]), but only centralized algorithms have been proposed.
We now show that the load-spillage characterization makes distributed optimal algorithms possible.
Fixing ρ ∈ [0, 1) and using the parameterization γ = γ(s, ρ)
in Section III-A, we can view the utility maximization problem
as an optimization over the load vector s. The maximization
problem can then be solved by application of any general ascent
method [17] in the feasibility region of s. We obtain an ascent
direction that can be calculated locally by each MS, and employ
the mechanism in Algorithm 1 to update the SIR associated with
the links.
Consider the following load-update ∆s to the load vector s,
∆si =
Ui′ (γi )γi
− si , ∀i.
qi
(36)
We first present an iterative, distributed algorithm based on this
locally computable update, and then prove that (36) indeed represents an ascent direction and leads to an iteration whose fixed
point is the optimal load distribution problem (35). The intuition
behind the proposed load update will be further explained after
the proof of convergence is presented.
Algorithm 3—Utility Maximization over Bρ :
Parameters: step size δ > 0 and utility functions {Ui (γi )}.
Initialize: Arbitrary s[0] ≻ 0.
P
1) BS k broadcasts the BS-load factor ℓk [t] = i∈Sk si [t].
2) Compute the spillage-factor ri [t] according to (17).
3) Assign SIR values γi [t] = ρsi [t]/ri [t].
4) BS measures the resulting interference qi [t].
5) Update the load factor si [t] in the ascent direction given
by (36)
si [t + 1] = si [t] + δ∆si [t].
•
•
Continue: t := t + 1.
Remarks: The load update can be done either at the BS(BSControl) or at the MS(MS-Control).
1) The load update for MS-Control requires the calculation
of the resulting interference based on the assigned SIR
and the transmitted power, but otherwise results in a completely distributed implementation.
2) BS-Control requires a mechanism to have the MS calculate riext and convey that to the BS, but provides the flexibility of letting the BS decide on the utility function asso-
ciated with the links it is serving. BS-Control is still distributed in the sense that each BS in the network decides
on the SIR assignment of the links it is serving, with the
overall SIR assignment turning out to be optimal for the
entire multicellular network without the need for a centralized SIR assignment across the BSs in different cells.
3) The interference qi [t] can be measured after the power updates, e.g., by the Foschini-Miljanic method, converge.
We empirically find in our simulations that convergence
is achieved even when qi [t] is the interference resulting
from a single update in the power level (10), i.e., even
when power control is running at the same time-scale as
SIR assignment.
From Lemma 5, as ρ → 1, the following approximation becomes accurate:
1
q[t] ≈
q̃(γ[t]),
1−ρ
where q̃(γ[t]) is a suitably normalized positive right eigenvector of GD(γ[t]). Algorithm 4 converges to the globally utilityoptimal SIR assignment when ρ → 1. This includes the important, zero-noise case often considered in the research literature,
for which ρ = 1 is the feasibility region.
Theorem 3: For ρ → 1 and sufficiently small step size δ >
0, Algorithm 3 converges to the globally optimal solution of
problem (35).
Proof: Consider the sum utility that is being maximized
in (35) as a function of s, Fρ (s) = U (γ(s)). Let ∆s be the
update direction of Algorithm 3 as given in (36) with the load
update given by
s[t + 1] = s[t] + δ∆s[t].
(37)
We show in Appendix D that (37) is a general ascent update.
In addition, the Hessian’s norm is bounded since the first and
second derivatives of the utilities are bounded, the load s ≻ 0
at every iteration t (since zero SIR will not be assigned to any
MS due to assumption 3 on utility function), and the connection matrix G is primitive. This implies the Lipschitz continuity
property [18] on Fρ (s) and the existence of a non-zero step size
δ such that U (γ[t + 1]) > U (γ[t]). With Lipschitz property,
the descent lemma in [18] implies convergence of this ascent
algorithm.
If s∗ is the point of convergence of the algorithm, then the
corresponding spillage r∗ = GT s∗ and SIR {γ ∗ = ρs∗ /r∗ }
satisfy ∆s∗ = 0 so that
Ui′ (γi∗ ) = ri∗ q̃i∗ (γ ∗ ),
i = 1, . . . , M.
(38)
This is precisely the KKT condition for {γ ∗ , ν ∗ = 1} the optimization problem (35) as shown in Appendix E. Since the optimization problem (35) is convex after a change of variable on s,
any point that satisfies the KKT condition is indeed the optimal
solution to the problem [17].
Since Ui′ (γi ) is decreasing in γi , we see that larger the
spillage and larger the interference, the lower the optimal γi is.
This makes intuitive sense, since a link with large interference
requires a larger power allocated for a given SIR and results in
9
larger interference to neighboring links. Similarly, a link with
larger spillage easily carries over interference into neighboring
links and hence should be assigned a smaller SIR.
B. Utility Maximization with Power and Interference Constraints
The utility maximization problem over the power (Case P)
and interference (Case Q) constrained SIR region can be written
as follows
P
Case P : maximize
i Ui (γi )
(39)
subject to p(γ) pm
P
Case Q : maximize
i Ui (γi )
(40)
subject to q(γ) qm
The Lagrangian for this convex optimization problem can be
written as
Case P : Lp (γ, ν) = U (γ) + ν T (pm − p)
(41)
Case Q : Lq (γ, ν) = U (γ) + ν T (qm − q)
(42)
where {νi ≥ 0, i = 1, . . . , M } are the Lagrange multipliers.
Using standard convex optimization theory [17], we attempt
to move the SIR vector γ in a direction that increases the Lagrangian, thereby attempting to solve the weighted optimization
max L(γ, ν),
γ
problem (39) for the power constrained case and (40) for the
interference constrained case.
Proof: For a given ν, consider the Lagrangian as a function of the load:
Fp (s)
= Lp (γ(s, ν), ν)
(43)
Fq (s)
= Lq (γ(s, ν), ν)
(44)
The update direction ∆s of the algorithm is given in (36) with
the load update given by (37). We show in Appendix D that (37)
is an ascent update with respect to Fp (s) and Fq (s). In addition,
norms of the Hessians of Fp (s) and Fq (s) are bounded since
the first and second derivatives of the utilities are bounded, the
load s ≻ 0 at every iteration t, and the connection matrix G is
primitive. This implies the Lipschitz continuity property [18]
on Fp (s) and Fq (s) and the existence of a non-zero step size δ
such that L(γ[t + 1], ν) > L(γ[t], ν) where L represents Lp or
Lq . With Lipschitz property, the descent lemma in [18] implies
convergence to the Lagrangian maximum.
The update for the price vectors ν corresponds to the gradient for the Lagrange multipliers if we allow the convergence to
the Lagrangian maxima over s at every price ν[t]. Of course,
in practice, both the load factor s and price ν are updated simultaneously. Such a joint update can be proven to converge to
the optimum of the original constrained optimization problem if
the Lagrangian is concave and the maximized Lagrangian for a
given price is convex [19]. This is indeed true and therefore the
algorithm converges to the optimal.
Consider now the interference constrained case. If {s∗ , ν ∗ }
is a fixed point of the algorithm, then the corresponding spillage
r∗ = GT (s∗ + ν ∗ ) and SIR {γ ∗ = s∗ /r∗ } satisfy ∆s∗ = 0 so
that
where L = Lp or Lq depending on the problem. Similar to Section V-A, we approach the problem in the domain of the load
vectors s rather than the domain of SIR vectors γ, and use the
ascent direction ∆s in (36) for this purpose. The prices are optimized by gradient updates. The proposed algorithm is as folUi′ (γi∗ ) = ri∗ qi∗ (γ ∗ ),
i = 1, . . . , M.
(45)
lows:
∗ ∗
Algorithm 4—Utility Maximization over B(pm ) or B(qm ): This is precisely the KKT condition at {γ , ν = 1} for the
optimization problem (40) as shown in Appendix E. Since the
optimization problem (40) is convex after a change of variable,
any point that satisfies the KKT condition is indeed the optimal
• Parameters: Step sizes δ1 > 0 for load update, δ > 0 for
solution to the problem [17].
price update, utility functions Ui (γi ), and interference or
power constraints, pm or qm .
VI. S IMULATION R ESULTS
• Initialize: Arbitrary s[0] ≻ 0 and ν[0] 0.
To illustrate our distributed power control algorithms for
1) a) Case P: Compute spillage ri [t] according to (27)
utility-optimal
SIR assignment, we simulate the load control alb) Case Q: Compute spillage ri [t] according to (28)
gorithm
in
a
cellular
network, using a simplified version of the
2) a) Case P: Assign γi [t] according to (27)
realistic
3GPP
uplink
evaluation model [20] that is adopted by
b) Case Q: Assign γi [t] according to (28)
the
wireless
industry.
3) Measure resulting interference qi [t]
The model consists of 19 cells arranged in a three ring hexag4) Update load si [t] in the ascent direction given by (36)
onal structure as shown in Fig. 3. Each cell is divided into three
si [t + 1] = si [t] + δ1 ∆si [t].
identical 120 degree sectors for a total of 57 base station sectors. The BS angular antenna sectorization pattern is based on
5) a) Case P: Update price νi [t] according to (29)
the commercial Decibel DB 932DG65T2E antenna, which has
b) Case Q: Update price νi [t] according to (30)
a 65 degree 3 dB bandwidth, 15 dB antenna gain, and 20 dB
Continue: t := t + 1
front-to-back rejection. This antenna, used in the PCS 1.9 GHz
band, is typical of commercial antennae used in cellular wireTheorem 4: For sufficiently small step size δ1 > 0 and δ > less networks. The mobiles use single omni-directional anten0, Algorithm 4 converges to the globally optimal solution of nae. We adopt the path loss model with a path loss exponent of
10
TABLE II
4
S ECTOR CAPACITY AND FAIRNESS UNDER VARIOUS UTILITY FUNCTIONS .
3
Utility function
2
1
Pseudo-Linear
α = 1 (Log)
α=2
α=3
0
−1
Sector capacity
(bps / Hz / sector)
1.77
1.76
1.56
1.45
10% User capacity (bps / Hz)
0.054
0.057
0.076
0.086
−2
−3
−4
−5
−4
−3
−2
−1
0
1
2
3
4
5
Fig. 3. Hexagonal cellular network for the uplink simulation consisting of 19
cells with wrap around. Cells are three-way sectorized for a total of 57 base station sectors. Path loss accounts for log-normal shadowing, and angular antenna
pattern. Plotted are the location of 300 random mobiles connected to the lower
sector of the center cell. Cell radii are normalized to unity.
3.7 and log-normal shadowing of 8.9 dB. To avoid edge effects,
the sector boundaries at the edge of the hexagonal structure are
“wrapped around” to the opposite cell. Mobiles are distributed
uniformly in the 19 cell region and then connect to the base station sector to which the path loss is minimum. Fig. 3 shows the
location of 300 random mobiles connected to the lower sector
of the middle cell under this path model.
In all the simulations below, each realization of the network
consists of 10 randomly selected mobiles in each base station
sector, for a total of 570 mobiles. As a measure of the QoS, we
will use the capacity formula (31), where we assume that each
mobile is allocated equal uplink bandwidth fraction, wi /W =
1/10 where W is the total bandwidth. To represent the sum
utility in a meaningful unit, we plot the inverse of the average
of utilities of QoS factors, which is a quantity whose units are
the same as the QoS factor and is denoted as “QoS factor” on
the y-axis in Figures 4 and 5. For example, the QoS factor thus
computed for logarithmic utility is the geometric mean of the
user QoS factors:
"
#
M
M
Y
1
1 X
exp
U (γ) = exp
log βi (γi ) =
βi (γi )1/M .
M
M i=1
i=1
We first illustrate the convergence of the distributed algorithm, Algorithm 3, using the logarithmic utility defined in Section IV. Fig. 4 shows the convergence of the distributed algorithm as a function of the iteration number for a particular distribution of the MSs. The algorithm is initialized with a random
positive load vector s, and the step size is taken as δ1 = 0.1. The
ROT limit, −10 log(1 − ρ) is set to 10 dB. Also plotted is the
utility which was numerically attained by centralized computation. It can be seen that the utility from the distributed algorithm
is almost identical to the maximal utility within about 30 iterations.
In Fig. 5, we plot the convergence of the algorithm with various utility functions. As expected, the pseudo-linear utility con-
verges to the maximum QoS value while the α-fair utility with
α = 3 converges to the minimum. The difference between the
pseudo-linear and logarithmic utility is minimal since this experiment is run under heavy loaded scenario.
Fig. 6 shows the cumulative distribution of the user capacities
after the final iteration of Algorithm 4. As discussed earlier, one
of the important features of the utility approach is that this QoS
distribution can be varied by appropriately adjusting the utility.
Fig. 6 shows the resulting QoS distribution for the pseudo-linear
utility, log utility and the α-fair utilities with α = 2 and α = 3,
all with the same realization of MS distribution in the network.
It can be seen that the variance in the mobile’s QoS with the
{2, 3}-fair distributions is significantly smaller than with the
logarithmic and pseudo-linear utility functions. Consequently,
fewer mobiles are seen with either very high or very low achievable capacities.
As is widely known in the research literature and discussed
in Section IV, the coefficient α in the α-fair exponential distribution can be seen as a fairness parameter, with larger α resulting in greater equalization of allocated QoS. However, in
general, while increasing fairness improves the performance of
worst case mobiles, it tends to reduce the average sector capacity. To illustrate this fairness-efficiency tradeoff in wireless cellular networks, Table II shows the average sector capacity and
performance of the 10% worst mobiles under the various utility
functions. It can be seen that average sector capacity drops from
1.77 to 1.46 bps/Hz/sector when we go from pseudo-linear utility to α-fair utility with α = 3. However, the QoS for the 10%
worst mobile increases from 0.054 to 0.086 bps/Hz. By appropriately adjusting the utility function, the network operator can
tradeoff fairness and total sector capacity.
Finally, we simulated Algorithm 4 with interference constraints. We assumed that the noise vector ηi was constant for all
mobiles, and the interference constraint was taken as qm = κη,
where κ > 1 is the ROT limit. We used the logarithmic utility,
and ran Algorithm 4 with step sizes δ1 = 0.1 and δ = 0.01.
Fig. 7 shows the convergence of the algorithm for various desired ROT levels for logarithmic utility function. Plotted is the
target ROT and the actual maximum ROT over the 57 BS sectors
after 25 iterations of the algorithm. The algorithm converges to
within a fraction of a dB of the target within the 25 iterations.
Fig. 8 plots the convergence of the interference over time for the
logarithmic utility and the 2-fair utility at a ROT level of 10dB.
The convergence is slower for the fairer utility function but still,
we notice convergence within 40 iterations. Fig. 9 shows the
11
1
0.105
0.12
0.9
0.1
0.8
Cumm Probability
0.095
QoS (bps/Hz/user)
QoS (bps/Hz/user)
0.1
0.09
Centralized Gradient Control
Distributed control
0.085
0.08
0.06
pseudo−linear
logarithmic
2−fair
3−fair
0.04
0.7
0.6
0.5
0.4
pseudo−linear
logarithmic
2−fair
3−fair
0.3
0.2
0.08
0.1
0.02
0.075
0
0
5
10
15
20
25
30
35
40
45
50
0
Iteration
0
10
20
30
40
50
60
70
80
90
Iteration
Fig. 4. Convergence of Algorithm 4 with the
logarithmic utility. The simulation is based on
one realization of the cellular network in Fig. 3
with 10 mobiles per sector for a total of 570 mobiles. The logarithmic utility is plotted as the geometric mean user capacity.
Actual ROT after 25 iterations
Target ROT
8
6
4
2
0
0
12
10
8
6
4
5
10
Target Rise Over Thermal (dB)
Fig. 7.
Convergence with interference constraints expressed as ROT limits. Plotted is the
desired and actual maximum ROT level.
15
2
0.2
0.25
0.3
0.35
0.4
0.45
0.5
1.8
Sector capacity (bps/Hz/cell)
10
0.15
2
Limit
logUtil, max Interference
logUtil, mean Interference
2−fair, max Interference
2−fair, mean Interference
14
12
0.1
User capacity (bps/Hz/user)
16
Raise Over Thermal(dB)
Actual Rise Over Thermal (dB)
14
0.05
Fig. 6. User capacity distribution after 50 iterations of the distributed algorithm under different utility functions: α-fair distribution with
α = 1, 2, 3, and the pseudo-linear utility. The
final distribution is based on one realization of
randomly located mobiles, with 10 mobiles per
sector.
Fig. 5. Convergence of utility functions for different utilities. Pseudo-linear utility functions attains the maximum utility and the 3-fair utility
attains the minimum.
16
0
100
1.6
1.4
1.2
1
0.8
0
10
20
30
40
50
60
70
Iteration
Fig. 8.
Convergence of the interference expressed as ROT for different utilities.
mean sector capacity achieved with the distributed algorithm as
a function of the ROT limit, κ. The graph quantifies the intuition that raising the ROT limit increases the maximum spectral
efficiency, with a diminishing marginal impact. In this way, the
network can be configured to tradeoff ROT level (which affects
the absolute power usage of the mobiles) and the overall spectral
efficiency.
We see from these simulations that, in a practical cellular
model with a large number of base station sectors, the distributed power control algorithms are stable and converge to the
globally optimal SIR solution within a small number of iterations. Moreover, the algorithms can be applied with a general
utility function. This flexibility provides network operator several adjustable knobs to trade-off various considerations such
as fairness, total throughput, loading effects, differentiated user
QoS and ROT levels.
80
0.6
0
5
10
Rise over thermal (dB)
15
Fig. 9. Mean sector capacity as a function of
ROT limit.
VII. C ONCLUSION
We solved the following problem in this paper: distributed
uplink power control over multiple cells for globally optimal
SIR assignment. This covers a variety of earlier results since
the early 1990s as special cases, including power control methods for fixed target SIR, for suboptimal Nash equilibrium in SIR
optimization, and for optimal SIR assignment obtained through
centralized computation. The first key step is the load-spillage
characterization (which provides a complementary angle to the
traditional power-interference view on the feasible SIR region
and its Pareto-optimal boundary), and the second key step is the
locally computable ascent direction to “slide” along the Paretooptimal SIR boundary. We also quantify the intuition that a
jointly optimal SIR assignment and power control should take
into consideration both spillage and interference for each link,
with the spillage being a measure of the emitted interference.
12
This characterization enables us to decouple the optimization
problem with complex coupling across all MS in many cells.
This work clears the bottleneck of “global coupling” towards solving the general resource allocation problem for multicellular uplink in wireless cellular networks. The algorithms
can be readily implemented in commercial systems such as the
Flash-OFDM system [21], and the convergence, optimality, and
efficiency-fairness tradeoff properties of the algorithms verified
through 3GPP evaluation tools.
The second bottleneck of “non-convexity” remains a difficult
challenge, especially when antenna beamforming, base station
assignment (due to mobility), and opportunistic scheduling (under fast fading) are performed jointly with power control for
optimal SIR assignment. Some of these additional degrees of
freedom have recently been incorporated [13] into the framework of distributed solution developed in this paper.
be increased, the primitivity assumption on G implies that the
network is fully connected, so pi has to increase to maintain the
same γi . This is impossible, thus proving the claim by contradiction. The proof for B(qm ) follows similarly.
A PPENDIX C: P ROOF
GD(γ(s, ρ))q̃ = ρGD(γ 1 )q̃ = ρq̃,
so q̃ is a positive right eigenvector of GD(γ(s, ρ)).
Part (b) follows from the technique used in the expansion of (I − GD(γ))−1 in [15]. From [15], we can write
−1
(I − GD(γ)) as
n
We state the following results on gradients, skipping the proof
due to space limitation. If ρ(D(γ)G) is the spectral radius
D(γ)G, then
(46)
where D(ei ) has zero everywhere, except the ith diagonal element which is 1, r is the left eigenvector of D(γ)G, and q̃ is
the right eigenvector of GD(γ). If p(γ) and q(γ) are given
according to (8) and (9), then
∂p
∂γ
∂q
∂γ
(47)
= (I − GD(γ))−1 GD(q).
(48)
If γ = γ(s, ρ) for some fixed ρ, then
∂γ
= D(1/r)(ρI − GD(γ))T .
∂s
(49)
−1
=
m
k
XX
1
(j − 1)!
Z1,1 +
j Zk,j (51)
1−ρ
(1
− λk )
j=1
k=2
where n is the number of distinct eigen-values, mk is the algebraic multiplicity of eigen value λk and Zk,j are known as the
principal component matrices. Further, we have from [15] that
Z1,1 = q̃sT /(sT η)
−1
Since qρ = (I − GD(γ))
(52)
η, Lemma 5(b) follows.
OF
A SCENT D IRECTION
Consider the functions
Fρ (s) = U (γ),
T
Fq (s) = U (γ) + ν (qm − q(γ)),
sT GD(γ) = ρsT
(53)
T
T
(s + ν )GD(γ) = sT
Fp (s) = U (γ) + ν T (pm − p(γ)),
(sT G + ν T )D(γ) = sT ,
based on which we define
Jρ (s) = ∇F Tρ ∆s, Jq (s) = ∇F Tq ∆s, Jp (s) = ∇F Tp ∆s (54)
If γ = γ q (s, ν) for some fixed vector ν, then
∂γ
= D(1/r)(I − GD(γ))T .
∂s
(I − GD(γ))
A PPENDIX D: P ROOF
= (I − D(γ)G)−1 D(q)
L EMMA 5
For part (a), let γ 1 = γ(s, 1) and let q̃ be a positive right
eigenvector with GD(γ 1 )q̃ = q̃. We can normalize q̃ so that
sT q̃ = sT η. For ρ < 1, (12) shows that γ(s, ρ) = ργ 1 , hence
D(γ(s, ρ)) = ρD(γ 1 ). Consequently,
A PPENDIX A: G RADIENT E VALUATION
∂ρ
= rT D(ei )q̃
∂γi
OF
(50)
A PPENDIX B: PARETO -O PTIMALITY P ROOF
We first consider ∂Bρ . From (46), we have ∂ρ/∂γi = ri q̃i .
Since GD(γ) and D(γ)G are primitive, and primitive matrices have positive left and right eigenvectors associated with the
simple eigenvalue ρ, ri > 0, q̃i > 0, we have ∂ρ/∂γi > 0. This
implies that the Pareto-optimal boundary of Bρ is indeed given
by ρ(GD(γ)) = ρ.
Next, we prove the result for B(pm ). Suppose γ is a SIR assignment that satisfies the Pareto-optimal constraint given in the
theorem but is not Pareto-optimal, then γi cannot be increased
without decreasing any other SIR assignment since pi is already
at its maximum. So let γj , j 6= i be a SIR assignment that can
where ∆s is given by (36). Let F (s) represent any one of the
functions defined in (53) and J(s) represent the corresponding
function in (54). We need to prove that ∆s is an ascent direction
for F (s). To prove this, we need to show that for all s > 0,
J(s) ≥ 0 with J(s) = 0 if and only if ∇F (s) = 0. To this
end, we first state the following Lemma, the proof of which is
omitted due to space limitation.
Lemma 6—Positivity Lemma: Suppose T is a matrix with
positive entries, i.e. Tij > 0 for all i and j.
(a) Suppose that q > 0 and s > 0 are vectors satisfying,
Tq = ρq, TT s = ρs,
(55)
for some eigenvalue ρ > 0. Define the function,
H(x) = xT D(s)(ρI − T)D(q)x.
(56)
13
Then, H(x) ≥ 0 for all vectors x, with H(x) = 0 if and
only if x = β1 for some scalar β.
(b) Suppose that q > 0 and s > 0 are vectors satisfying,
Tq < ρq, TT s < ρs,
T
=
=
Ui′ (γi ) = ri qi .
(57)
for some ρ > 0. Then, H(x) ≥ 0 for all vectors x, with
H(x) = 0 if and only if x = 0.
Continuing with the proof, we first consider F = Fq and set
T = GD(γ). We now evaluate ∂F (γ)/∂γ. We let U ′ (γ) =
∂U/∂γ and use result from (48) to derive that
(∂F/∂γ)
Therefore, s and q satisfy (57) in Lemma 6(b). Hence, J(s) ≥ 0
for all s. Moreover, J(s) = 0 if and only if x = 0. But, using
(61), x = 0 if and only if,
∂q
∂γ
U ′ (γ)T − ν T (I − T)−1 GD(q). (58)
U ′ (γ)T − ν T
From (60), this condition is equivalent to ∂F/∂γ = 0. We
conclude that
J(s) = ∇F T ∆s ≥ 0,
with equality if and only if ∇F (s) = 0. A similar computation
proves the result for F = Fp .
For F = Fρ ,
T
= U ′ (γ)T
(∂F/∂γ)
(63)
Now combining (63) with (49),
Now
r = GT (s + ν) = GT (D(γ)r + ν),
∂γ
∂s
= (U ′ (γ)D(1/r)(ρI − T)T
T
∇F T (s) = (∂F/∂γ)
and consequently,
r =
=
(I − GT D(γ))−1 GT ν
T
T −1
G (I − D(γ)G )
T
T −1
ν = G (I − T )
ν.
T
(59)
∂γ
∂s
= (U ′ (γ) − rD(q))T D(1/r)(I − T)T
U ′ (γ)
= (
− 1)T D(q)(I − T)T
rq
= xT D(q)(I − T)T ,
(60)
T
x=
U ′ (γ)
− 1.
rq
(61)
Also, since γi = si /ri , we can rewrite ∆si as
′
Ui′ (γi )γi
Ui (γi )
∆si =
− si = si
− 1 = si xi ,
qi
qi ri
(62)
J(s) = xT D(q)(I − T)T D(s)x.
Now observe that,
s = D(γ)r = D(γ)GT (s + ν) = TT (s + ν) > TT s,
Tq = q − η < q.
(66)
Substituting (64) and (66) into (54),
J(s) = xT D(q)(I − T)T D(s)x.
Since q and s are, respectively, right and left eigenvectors of T,
Lemma 6(a) shows that J(s) ≥ 0 for all s. Moreover, J(s) = 0
if and only if x = β1 for some scalar constant β. which implies
that xT D(q)(ρI − T)T = βqT (ρI − T)T = 0. From (64), this
condition is equivalent to ∇F (s) = 0. We conclude that
J(s) = ∇F T ∆s ≥ 0,
A PPENDIX E: KKT C ONDITION
Substituting (60) and (62) into (54),
and note that
Therefore, we have
with equality if and only if ∇F (s) = 0.
or in vector notation,
∆s = D(s)x.
(65)
∆s = D(s)x − s.
= (∂F/∂γ)
where
U ′ (γ)
.
rq
x=
Now combining (59) with (50), we have
∇F T (s)
(64)
where
Substituting this identity into (58), we see that
(∂F/∂γ) = U ′ (γ)T − rT D(q).
= xT D(q)(ρI − T)T ,
We derive the KKT condition [17] for the utility optimization
problems. The Lagrangian for the optimization problem (35) is
given by
L(γ, ν) = U (γ) − ν(ρ(D(γ)G) − ρ)
(67)
with ν ≥ 0 being the Lagrange multiplier. The KKT condition
is given by ∇L = 0 which implies that ∇U = ν∇ρ(D(γ)G).
From (46), this can be written as
Ui′ (γi ) = νri (γ)q̃i (γ)
(68)
where r is the left eigenvector of D(γ)G and q̃ is the right
eigenvector of GD(γ).
14
The Lagrangians for the power and interference constrained
cases are given in (41) and (42) respectively. The KKT condition can be written as ∇U = ν T ∇(p), which using (47)
and (48) respectively, we simplify to
Ui′ (γi ) = ri (γ)qi (γ)
(69)
where q is given by (9) and r = rp as given by (25) for the
power constrained case and r = rq as given by (26) for the
interference constrained case.
ACKNOWLEDGMENTS
The authors would like to acknowledge very helpful discussions with Arnab Das at Qualcomm Flarion Technologies.
R EFERENCES
[1] P. Hande, S. Rangan, M. Chiang, ”Distributed Uplink Power Control for
Optimal SIR Assignment in Cellular Data Networks”, IEEE INFOCOM
2006, April 2006.
[2] J. Zander, ”Performance of Optimum Transmitter Power Control in Cellular
Radio Systems”, IEEE Transactions on Vehicular Technology, VT-41, no.
1, pp. 57-62, February 1992.
[3] G. J. Foschini and Z. Miljanic, ”A Simple Distributed Autonomous Power
Control Algorithm and it’s Convergence”, IEEE Transactions on Vehicular
Technology, VT-42, no. 4, pp. 641-646, April 1993.
[4] R. D. Yates, ”A Framework for Uplink Power Control in Cellular Radio
Systems”, IEEE Journal on Selected Areas in Communications, Vol. 13,
no. 7, pp. 1341-1347, September 1995.
[5] N. Bambos, S. C. Chen, G. J. Pottie, ”Radio Link Admission Algorithms
for Wireless Networks with Power Control and Active Link Quality Protection”, Proc. IEEE INFOCOM 1995: 97-104
[6] D. Mitra and A. Morrison, “A distributed power control algorithm for bursty
transmissions in cellular, spread spectrum wireless networks”, Proc. 5th
WINLAB Workshop, Rutgers University, New Brunswick, NJ, 1995.
[7] C. Saraydar, N. Mandayam, and D. Goodman, “Pricing and power control
in a multicell wireless data network”, IEEE Journal on Selected Areas in
Communications, vol. 19, no. 10, pp. 1883-1892, October 2001.
[8] T. Javidi, ”Decentralized Rate Assignments in a Multi-Sector CDMA Network”, IEEE Globecom Conference, December 2003.
[9] M. Chiang and J. Bell, “Balancing supply and demand of wireless bandwidth: Joint rate allocation and power control”, Proc. IEEE INFOCOM,
Hong Kong, China, March 2004.
[10] D. O’Neill, D. Julian, and S. Boyd, “Adaptive management of network
resources”, Proc. IEEE VTC, October 2003.
[11] M. Xiao,N. B. Shroff, E. Chong, “A utility-based power-control scheme in
wireless cellular systems”, IEEE/ACM Transactions on Networking, Volume 11, Issue 2, pp. 210 - 221, April 2003.
[12] H. Boche and S. Stanczak,“Convexity of some feasible QoS regions and
asymptotic behavior of the minimum total power in CDMA systems”, IEEE
Transactions on Communications, Volume 52, Issue 12, pp. 2190 - 2197,
December 2004.
[13] T. Lan, P. Hande, and M. Chiang, “Joint uplink power control, bandwidth allocation, and beamforming for optimal QoS in wireless networks”,
Preprint, 2006.
[14] J. Mo and J. Walrand, “Fair end-to-end window-based congestion control”, IEEE/ACM Transactions on Networking, Volume 8, Issue 5, pp. 556
- 567, October 2000.
[15] E. Deutsch and M. Neumann, “Derivatives of the Perron root at an essentially nonnegative matrix and the group inverse of an M-matrix”, Journal of
Mathematical Analysis and Applications, vol. I-29, no. 102, 1984.
[16] R. A. Horn, C. R. Johnson, Topics in Matrix Analysis, Cambridge University Press, Cambridge, 1991.
[17] S. Boyd, L. Vandenberghe, Convex Optimization, Cambridge University
Press, 2004.
[18] D. P. Bertsekas, Nonlinear Programming, 2nd Ed., Athena Scientific,
1999.
[19] R. T. Rockafellar, Saddle-points and convex analysis, Differential Games
and Related Topics, H. W. Kuhn and G. P. Szego, Eds. North-Holland, 1971.
[20] 3GPP TR 25.896 (2004-02), Feasibility Study for Enhanced Uplink for
UTRA FDD, Document R1-040346, V1.3.0.
[21] Flash-OFDM, http://www.qualcomm.com/qft/
PLACE
PHOTO
HERE
Prashanth Hande Prashanth Hande is a Graduate student at Princeton University and a Staff Engineer at
Qualcomm Flarion Technologies. He received the
B.Tech in Electrical Engineering at the Indian Institute of Technology, Mumbai in 1998 and his M.S in
Electrical Engineering at Cornell University in 2000.
Prashanth Hande’s interests are in wireless communications, network architectures and boradband access
networks.
Sundeep Rangan Dr. Sundeep Rangan is a Director
of Engineering at Qualcomm Flarion Technologies.
He received the B.A.Sc. in Electrical Engineering at
the University of Waterloo, Canada in 1992 and his
M.S and Ph.D. in Electrical Engineering at the UniPLACE
versity of California, Berkeley in 1995 and 1997, rePHOTO
spectively. After a postdoctoral research fellowship at
HERE
the University of Michigan, Ann Arbor, he joined Bell
Labs, Lucent Technologies in 1998. At Bell Labs, Dr.
Rangan worked on the development of Flash-OFDM,
a novel air interface for broadband, cellular wireless.
The Bell Labs department was spun off to form a separate company, Flarion
Technologies, in October 2000. Dr. Rangan was one of four founding members. Flarion was acquired by Qualcomm in Jan 2006, and Dr. Rangan remains
involved in physical layer design and implementation. Dr. Rangan’s interests
are in wireless communications, information theory, control theory and signal
processing.
PLACE
PHOTO
HERE
Mung Chiang Mung Chiang (S’00 – M’03) received
the B.S. (Hon.) degree in electrical engineering and
in mathematics, and the M.S. and Ph.D. degrees in
electrical engineering from Stanford University, Stanford, CA, in 1999, 2000, and 2003, respectively. He
is an Assistant Professor of Electrical Engineering,
and an affiliated faculty of Applied and Computational
Mathematics at Princeton University. He conducts research in the areas of optimization of communication
systems, theoretical foundations of network architectures, algorithms in broadband access networks, and
information theory.
Professor Chiang has been awarded a Hertz Foundation Fellowship and received the Stanford University School of Engineering Terman Award for Academic Excellence, the SBC Communications New Technology Introduction
contribution award, the National Science Foundation CAREER Award, and the
Princeton University Howard B. Wentz Junior Faculty Award. He is the Lead
Guest Editor of the IEEE J. Sel. Area Comm. special issue on Nonlinear Optimization of Communication Systems, a Guest Editor of the IEEE Trans. Inform.
Theory and IEEE/ACM Trans. Networking joint special issue on Networking
and Information Theory, and the Program Co-Chair of the 38th Conference on
Information Sciences and Systems.
Xinzhou Wu Xinzhou Wu was born in Jiangyin,
China, on September 23, 1975. He attended Tsinghua University, Beijing, China, where he received
the bachelor of engineering degree in electronics engineering in 1998. He received his M.S. in 2000 and
PLACE
Ph.D. in 2004, both in electrical engineering from the
PHOTO
University of Illinois at Urbana-Champaign. For the
HERE
academic year 2003-2004, he was a recipient of the
Vodaphone Fellowship. He was a member of technical staff at Flarion technologies from 2005 to 2006 and
he is currently a senior engineer at Qualcomm Flarion
technologies. His current research interests are in multiple antenna systems,
resource allocation in cellular communication systems, and information theory